swe-gym-pipeline

name	swe-gym-pipeline
description	Run the SWE-Gym training pipeline — skill learning, SFT, GRPO, or the full pipeline. Use when the user wants to train skills, fine-tune models, generate tasks, or run the gym pipeline.
allowed-tools	Bash(make *), Bash(./scripts/*), Bash(uv run *), Bash(torchrun *), Bash(docker *), Bash(ls *), Bash(cat *), Bash(grep *), Bash(test *), Read, Grep, Glob
argument-hint	[smoke-test|build-env|generate-tasks|validate-tasks|train|sft|grpo|evaluate|deploy-skills|resume|status|full]

SWE-Gym Training Pipeline

Orchestrates the full SWE-Gym pipeline: task generation, skill optimization, SFT/GRPO fine-tuning, and deployment.

Arguments: $ARGUMENTS

Pre-Flight Checks

Before any step, verify:

1. Docker is running: docker ps
2. .env exists with OPENAI_API_KEY (for GEPA skill training)
3. repo.conf has REPO_OWNER and REPO_NAME set

Pipeline Steps

Based on the argument ($ARGUMENTS), run the appropriate step:

smoke-test — Quick validation (~5 min, ~$0.50)

make smoke-test

Runs 3 tasks to verify the pipeline works end-to-end.

build-env — Step 1: Docker environment (~10-30 min)

make build-env

Builds Docker image with target repo + deps. Verify: docker images | grep swesmith

generate-tasks — Step 2: Synthesize bugs (~20-60 min)

make generate-tasks

Creates bug tasks from commit history. Check: ls logs/tasks/

validate-tasks — Step 3: Filter tasks (~30-60 min)

make validate-tasks

Runs test harness to keep only valid tasks. Check: ls logs/validated_tasks/

train — Step 4: GEPA skill optimization (~2-12 hrs, ~$30-80)

make train

Runs the GEPA gskill optimization loop. Monitor: tail -f gepa_results/logs/run_*/terminal.log

evaluate — Step 5: Compare baseline vs skills

make evaluate

deploy-skills — Step 6: Deploy learned skills

make deploy-skills

Copies best_skills.txt to .claude/skills/REPO_NAME/SKILL.md

sft — Fine-tune with SFT (after steps 6-7)

Requires trajectory data in logs/sft_data/. Run steps 6-7 first if missing:

make gen-trajectories
make prepare-sft

Then train (default Qwen3-8B LoRA):

make train-sft-unsloth

Or with a specific recipe:

uv run python scripts/08_train_sft_unsloth.py --recipe configs/unsloth/qwen3_8b_lora.yaml --four-bit

See configs/unsloth/ for all available recipes. Check make help for targets.

grpo — GRPO RL training (after step 3)

Requires validated tasks. The model learns by generating patches and getting rewarded:

make train-grpo

Or with specific recipe/options:

uv run python scripts/09_train_grpo.py --recipe configs/unsloth/grpo_qwen3_8b.yaml --reward-mode hybrid

Monitor rewards — expect 300+ steps before they start increasing.

hpo — Hyperparameter optimization

make hpo                    # 10 trials, ~30 min
make hpo-thorough           # 30 trials, ~2 hrs

Outputs hpo_output/best_recipe.yaml for use with SFT training.

resume — Continue interrupted training

make resume

status — Check current run

Show the latest run directory, logs, and results:

RUN_DIR=$(ls -td gepa_results/logs/run_* 2>/dev/null | head -1)
echo "Latest run: $RUN_DIR"
[ -n "$RUN_DIR" ] && tail -20 "$RUN_DIR/terminal.log"

full or empty — Run complete pipeline

Run steps 1-6 sequentially:

make build-env && make generate-tasks && make validate-tasks && make train && make evaluate && make deploy-skills

This takes 3-15 hours depending on repo size.

After Each Step

Always verify the step succeeded by checking for expected outputs before proceeding to the next step. If a step fails, check logs and report the error.

Key Files

• repo.conf — Target repo configuration
• .env — API keys
• configs/unsloth/ — Training recipe YAML files
• logs/ — Pipeline logs (tasks, trajectories, SFT data)
• gepa_results/ — GEPA training results
• sft_output/ — SFT model output
• grpo_output/ — GRPO model output